Is America a Tech Oligarchy or a Democracy?

Americans vote for the people who govern them. They do not vote for the people who own the cloud servers federal agencies run on, the platforms elections are argued on, or the AI models departments are now buying. If those two groups want different things, the question of who governs has two candidates, and only one of them stands for election.

The record of the last three years supports neither of the obvious answers. America's richest technology owners do not rule. They win some fights and lose others, including elections they pay for and quarrels with the White House, and when a substitute exists they can be replaced as suppliers. But the country does not govern its technology the way a democracy is supposed to govern anything, through laws an elected legislature writes and courts enforce. Congress has written few durable rules for these systems, so the terms are set in bargains between one elected office and a handful of owners, each holding levers over the other. Voters choose who sits on one side of that table. They have little say in what is agreed across it.

What the two words mean

Democracy is more than voting. The V-Dem Institute says electoral democracy exists when elections are free, fair and recurring, suffrage is universal, parties and candidates can form and compete, the contest is reasonably level in speech, media and civil society, and "elected officials wield political power de facto".¹ Its fuller standard, liberal democracy, adds checks on the executive by the legislature and the courts, and a rule of law that protects equal liberties.¹ The words "de facto" carry the whole question here. If elected officials hold authority on paper while the ability to act sits elsewhere, a country can pass every electoral test and still fail that clause.

Oligarchy is not the existence of rich people. Jeffrey Winters and Benjamin Page, using a material definition drawn from Aristotle, argued in 2009 that oligarchy "is not inconsistent with democracy" and that oligarchs need not hold office, conspire, or even engage much in politics to prevail.² On that definition America can be both at once, and the useful question becomes whether concentrated wealth turns into influence over consequential decisions that is disproportionate and durable.

Some private power is meant to sit beyond the voters' reach. In 2024 the Supreme Court wrote that a platform curating other people's posts is engaged in protected expression, and that "it is no job for government to decide what counts as the right balance of private expression".³ A liberal democracy protects that kind of power from majorities on purpose, so private power that voters cannot reach is not, by itself, a sign of oligarchy. What matters is whether such power is concentrated, and whether it turns into control over public decisions.

A test of who governs

Five questions follow from those definitions, covering elections, money, information, infrastructure and the checks on power. None yields a score, and each answer has a limit.

Do elections still change what the government does?

They do, and in technology policy they have done so abruptly. In October 2023 the president used the Defense Production Act to require companies developing the most powerful AI models to report to the government on an ongoing basis.⁴ On January 20, 2025, his successor revoked that order.⁵ In December 2025 the new president ordered the Justice Department to set up a task force whose "sole responsibility" is to challenge state AI laws that conflict with his policy, and asked for legislation that would preempt them.⁶ The federal government's stance toward the most powerful AI systems reversed with one election. V-Dem, which records steep decline elsewhere in American democracy, found the electoral components stable and judged the 2024 elections free and fair.¹

The limit is in what the elections changed. They changed executive orders, which the next election can change back, and they produced no law. Public majorities on technology have gone unlegislated for years. In 2023, 78 percent of Democrats and 68 percent of Republicans told Pew that there should be more government regulation of what companies can do with customers' personal information.⁷ In 2026 nearly two thirds of Americans, a majority in both parties, told the Annenberg Public Policy Center that government had done "too little" to regulate AI.⁸ By September 2026 Congress had 49 federal privacy proposals before it, had passed two narrow ones, and was still divided over whether a national law should override state laws and whether individuals could sue.⁹ The gap has more than one cause. The disputes over preemption and lawsuits divide legislators as well as lobbyists, and a system that rarely changes anything frustrates majorities whoever opposes them.

Does concentrated wealth outweigh the public?

The best-known evidence says it does. Martin Gilens and Benjamin Page tested 1,779 policy issues and concluded that economic elites and business groups "have substantial independent impacts on U.S. government policy, while average citizens and mass-based interest groups have little or no independent influence."¹⁰ It is also among the most contested findings in political science. Peter Enns showed that even where middle-income and affluent Americans disagree, policy tends to end up about where it would if policymakers had followed the middle.¹¹ Omar Bashir found 185 cases in the data where the two groups stood on opposite sides, and the median-income position won 47 percent of them; he also noted that the model explains less than 10 percent of the variation in outcomes.¹² Jeffrey Lax, Justin Phillips and Adam Zelizer, studying Senate votes from 2001 to 2015, found affluent influence "overstated and itself contingent on partisanship".¹³ Gilens replied that agreement is not influence: when ordinary citizens win, the original paper said, they may be "coincident beneficiaries rather than causes of the victory."¹⁴

The data in that debate are survey questions asked from 1981 to 2002.¹² That was before most of today's technology companies mattered in politics. Recent elections suggest that money buys presence more reliably than results.

Elon Musk spent $277 million in the 2024 cycle, more than any other donor to either party.¹⁵ In January 2025 he was put in control of the Department of Government Efficiency, created by executive order to modernize federal technology and software.¹⁶ In April 2025 he and groups tied to him spent more than $21 million on a Wisconsin Supreme Court race, and their candidate lost, trailing by nine points with 88 percent of the vote counted.¹⁷ In June 2026 a super PAC launched with more than $100 million in commitments from AI-industry figures and investors spent more than $8 million against Alex Bores, the author of New York's AI safety law, in a New York City congressional primary, while AI-safety groups spent more than $20 million for him.¹⁸ Bores lost, in what NOTUS called one of the industry super PAC network's first victories in a competitive election.¹⁹ The winner, Micah Lasher, campaigned on housing and affordability and had $10 million behind him from a super PAC backed by Michael Bloomberg.¹⁹ The industry got the result it paid for, though it was outspent and the winner had other backing. The voters chose from a menu that very rich people on every side had paid to shape.

Lobbying shows the same gap between access and results. Top technology and AI companies spent more than $100 million on lobbying in 2025, and Meta spent more than any other company in any industry.²⁰ In what its authors call the most extensive study of lobbying, covering 98 issues, 60 percent of campaigns failed to change policy, and resources explained less than 5 percent of the difference between winners and losers.²¹ Money pays most reliably in proximity to whoever wins. After control of Congress changed in 1994, companies whose boards were connected to the winning party gained federal procurement contracts, and those connected to the losing party lost them.²²

Who decides what the public sees?

The capacity is not in doubt. In 2010 Facebook showed a get-out-the-vote message to 61 million users, and researchers estimated that it produced about 340,000 additional votes, directly and through friends, roughly 0.14 percent of the voting-age population.²³ Control over what people believe is another matter. In a randomized experiment during the 2020 campaign, moving consenting Facebook and Instagram users to chronological feeds for three months substantially changed what they saw and did not significantly change their polarization or political knowledge.²⁴ The power to shape exposure is measured; the power to steer an electorate's opinions has not been shown.

What is documented is private decisions about visibility with public consequences. In January 2021 Twitter permanently suspended the sitting president's account, citing the "risk of further incitement of violence".²⁵ On January 7, 2025, Meta ended its third-party fact-checking program in the United States and lifted restrictions on some topics "that are part of mainstream discourse".²⁶ Platforms can act against elected power and can adjust to it, and both are private choices that the NetChoice opinion largely protects.

Ownership is another matter, and there the voters' representatives did reach. In April 2024 Congress passed a law requiring TikTok to be sold by its Chinese parent, ByteDance, or be cut off from American app stores and hosting services, and on January 17, 2025, the Supreme Court rejected the challenge to it.²⁷ The law left it to the president to determine whether a sale qualified, and on January 20 the new president ordered the Justice Department not to enforce it for 75 days.²⁷ Enforcement was delayed until January 23, 2026, and the day before, ByteDance said it had finalized a joint venture, with Oracle, Silver Lake and the Abu Dhabi firm MGX holding 15 percent each, ByteDance keeping 19.9 percent, and TikTok's recommendation algorithm to be retrained and secured in Oracle's cloud.²⁸ A White House official said the American and Chinese governments had both signed off.²⁸ Congress decided that TikTok's ownership had to change; the year of delay and the choice of owners were settled by negotiation.

How concentrated are the systems, and can the government do without them?

In the second quarter of 2026, Amazon, Microsoft and Google held 28, 20 and 15 percent of the world's cloud infrastructure market, and the top three held 67 percent of public cloud.²⁹ A federal review board wrote in 2024 that cloud providers "have become one of our most important critical infrastructure industries", and that the cloud "creates enormous efficiencies and benefits" while its ubiquity makes it a high-value target.³⁰ The same board found that an intrusion into Microsoft's email service by an actor linked to China, which reached the mailbox of the Commerce Secretary and took about 60,000 State Department emails, was "preventable", the product of "a cascade of security failures at Microsoft", and that the company's security culture was "inadequate".³⁰ Its remedy was a set of recommendations.³⁰ The board was advisory, and on January 20, 2025, a memo ending the memberships of every advisory committee in the Department of Homeland Security also ended its investigation into a Chinese-linked hack of American telecom companies.³¹

A dependence becomes a veto only when a private party can stop a public decision and the government cannot route around it in the time the decision allows. The Highest Monopoly found that kind of power over military networks in wartime; the civilian record mostly shows dependence without a switch. The clearest recent case of a supplier trying to set terms for the state's use of its product ended with the supplier outside. When Anthropic sought assurances that its models would not be used for fully autonomous weapons or domestic mass surveillance, the Defense Secretary accused it of trying to "seize veto power over the operational decisions of the United States military", and since then one court has found one legal basis for the Pentagon's blacklisting of the company unlawful while another has upheld the other.³² Claude, the model that drafted these sentences, is Anthropic's. By the time Anthropic sued in March 2026, Pentagon officials said xAI's and OpenAI's models had been cleared for use in classified systems.³³ A supplier with competitors holds a position it can lose, and no veto.

Regulation works more slowly and less completely. In a case filed in 2020, a federal judge found in August 2024 that Google had illegally maintained monopolies in search, and in September 2025 he declined to order the sale of its Chrome browser.³⁴ In September 2026 another judge declined to make Google sell its advertising exchange, the third time in a row a court had refused to break up a large technology company.³⁵ In the Meta case the judge found that the social media market had shifted drastically since 2020, and the search judge cited rising competition from generative AI companies such as OpenAI.³⁵ In two of those three cases, the market changed while the case was being heard. In those markets concentration bred its own challengers, and the courts counted on them. Fast-growing newcomers such as CoreWeave are appearing in cloud too, though the top three still hold 67 percent of public cloud.²⁹ Counting on challengers is a good argument for markets where one can appear within a few years, and a weaker one for the layers the government cannot do without in the meantime.

Legislatures have pushed back as well. In July 2025 the Senate voted 99 to 1 to strip a moratorium on state AI laws from its reconciliation bill, and Ted Cruz, who had proposed it, said its defeat was due in part to "outside interests".³⁶ In September 2025 California's governor signed what he called "first-in-the-nation frontier AI safety legislation".³⁷ New York's governor signed Bores's bill, which requires AI firms to publish their safety protocols and report incidents, late that year.¹⁸

Do the checks hold, and can the public shape the design?

The checks that matter here are the ones on the executive, because that is where the bargaining happens. V-Dem found that in 2025 legislative constraints on the executive lost a third of their value and fell to their lowest point in over 100 years, and that the United States lost its status as a liberal democracy for the first time in over 50 years.¹ It attributes the decline to executive overreach and pressure on the press, academia and dissent, and it singles out DOGE as a special case, citing estimates that more than 300,000 federal employees left government in 2025.¹ The richest technology owner's clearest mark on these measures was made from inside a government office. Freedom House cut its score for the United States from 84 to 81 out of 100, still rated Free.¹⁶ Both are expert judgments, and Andrew Little and Anne Meng have argued that objective indicators show little of the global decline that expert-coded indices report, and that coder bias may be part of the gap.³⁸

Public influence over the design of critical systems is the weakest part of the test, with one instructive exception. In February 2022, after an outcry from privacy advocates and lawmakers, the IRS said it would "transition away" from requiring facial recognition through the vendor ID.me for people who wanted to use its online services.³⁹ The next day ID.me announced it would drop the requirement for all its government clients, which included 10 federal agencies and 30 states.⁴⁰ The public reached the design because the design was in front of every taxpayer who logged in. Most design choices are out of sight: usage terms, data-sharing clauses, and the price at which a system first enters government. In August 2025 OpenAI offered its enterprise chatbot to every federal agency for $1 per agency for a year.⁴¹ A price like that can decide which systems the government's work is built on before anyone votes on anything.

Where the terms are set

The five answers share a pattern that neither label in the question describes. Where Congress has not legislated, the rules for privately owned systems are set by agreement between the executive branch and a few firms.

The pattern predates the current administration and has grown under it. In July 2023 the White House convened seven AI companies and announced safety and transparency commitments that the companies "have chosen to undertake".⁴² The reporting rule that followed was an executive order, and it fell with the next election. In August 2025 the government agreed to buy nearly 10 percent of Intel for $8.9 billion, most of it grants already awarded and not yet paid, and the president said the deal came out of a meeting with Intel's chief executive held days after he had called for that executive to resign.⁴³ Freedom House counted nine such ownership deals worth more than $10 billion by the end of 2025, and noted that some media and technology companies seeking regulatory approval for transactions made concessions or settled lawsuits brought by the president.¹⁶

The levers run both ways. In June 2025, during a public quarrel with Musk, the president wrote that the easiest way to save money was "to terminate Elon's Governmental Subsidies and Contracts"; Musk answered that SpaceX would "begin decommissioning its Dragon spacecraft immediately", and a little more than five hours later said it would not.⁴⁴ At the time SpaceX was the only American company able to carry crews to and from the space station.⁴⁵ The restoration of Ukraine's satellite links in October 2026 after a call between the same two men, told in The Highest Monopoly, had the same shape. People cross the same line. In June 2025 the Army commissioned the chief technology officers of Meta and Palantir and two current or former OpenAI executives as reserve lieutenant colonels, to serve without leaving their careers.⁴⁶

The record shows meetings, threats and agreements in sequence; it does not show any favor traded for any deal, and none should be read into it. What it does show is where decisions about these systems are made. The office holds contracts, licences, merger approvals and enforcement; the owners hold the services the government runs on, the platforms, the capital and the campaign money. Voters choose who holds the office's levers every four years. The legislature, which could fix the terms in law, has written few of them.

The sharpest instruments used on technology owners in 2025 and 2026 were discretionary: the threat to a contract, the blacklisting of a supplier, the terms of a sale. They work because they are fast and personal. That same discretion is what V-Dem and Freedom House find slipping out of the legislature's and the courts' control. A check on private power that answers to one office has the shape of the problem it is checking.

The label the evidence supports

The evidence does not support "technological oligarchy". On the definition that matters, wealth turning into durable control, the owners' record is mixed: they won in the antitrust courts and in one New York primary, and lost in Wisconsin, in the Senate's 99-to-1 vote, at the Pentagon and over TikTok, and Musk's own threat lasted five hours. Influence that wins some contests, loses others, and can be checked by a single office is not rule. Nor is America simply a democracy whose technology is controlled by oligarchs. Its government has compelled reporting from AI developers, forced the sale of a major platform, and replaced a supplier that tried to set terms.

Of the labels on offer, "plutocratic influence within democracy" comes closest, and it misses where the influence lands. The evidence supports something more specific: an electoral democracy with plutocratic influence, whose technology is governed by bargain rather than by law. Gilens titled his reply to one critic "The Insufficiency of 'Democracy by Coincidence'".¹⁴ On technology, the public gets what it wants when that coincides with what the two sides of the table agree.

What would have to change

Dependencies would have to be counted. No public register lists the functions the federal government cannot perform without one named company, and the Government Accountability Office and inspectors general are placed to keep one.

Second sources would have to be the rule wherever the government cannot do without a system. The Highest Monopoly showed Congress cutting a single-supplier satellite network and paying for several vendors instead; the same logic reaches the cloud the civilian government runs on and the AI models it is now adopting.

Deals would have to be published. Equity stakes, commitments made at White House meetings, and settlements reached while a company awaits approval are as consequential as contracts, and Congress could require them to be reported to it as contracts are.

Review bodies would have to outlast the officials they embarrass. A board that found the government's email supplier at fault lost all its members to a memo.

And the rules that matter would have to be written as law. An AI reporting rule imposed by one president and revoked by the next lasted less than fifteen months. A statute binds the next administration until Congress changes it. That is the step that would let voters reach the terms, and not only the people who negotiate them.